Multi-view field deep learning and data augmentation for heavy metals concentrations prediction in marine sediments based on visible and near-infrared spectroscopy.

Rapid and non-destructive monitoring of heavy metal pollution in marine sediments is very important for environmental protection and ecosystem health. However, the difficulties and high costs of marine sediment collection result in scarce data, which constrains the model's generalization. In this paper, a novel method fusing doubly regularized Wasserstein generative adversarial network (DR-WGAN-GP) and multi-view field long short-term memory (MVF-LSTM) was proposed. The MVF-LSTM method was designed that fused different view field spectral bands to build a multi-scale model for the heavy metal Cu and As concentrations in marine sediments. The DR-WGAN-GP method was used to generate high-quality, diverse sample data, effectively augmenting the original small scale training set. The results showed that MVF-LSTM had the highest accuracy in the Cu and As concentrations prediction model using the original samples. After augmenting the samples, DR-WGAN-GP combined with MVF-LSTM had the best prediction result. The R2, RMSE and RPD of Cu and As were 0.731 and 0.850, 2.674 and 1.228, 1.934 and 2.613, respectively. The combination of DR-WGAN-GP and MVF-LSTM was applied to the heavy metal prediction in marine sediment using Vis-NIR spectroscopy, providing a new technical way for deep learning modeling with limited sample in the field of environment.